Model evaluation for automated scoring of electropenetrography waveform data from mosquitoes.

Electropenetrography (EPG) is a powerful tool for quantifying how arthropods interact with their hosts, but manually labeling EPG feeding behavior waveforms is subjective, labor-intensive, and time-consuming. Machine-learning models for supervised classification have achieved widespread success in pattern recognition in scientific contexts and have the potential to automate and standardize waveform labeling. Here, the performance of various common machine learning methods for EPG waveform labeling was compared using a representative dataset of Culex tarsalis EPG waveforms; the top five models are presented. A neural network based on the UNet architecture with attention layers was the top-performing model, achieving an overall accuracy of 86% and a Macro F1 score of 0.78. To the authors' knowledge, this is the first attempt to apply machine learning to identify waveforms from a blood-feeding arthropod, rather than a plant-feeding hemipteran. Experiments indicate that previously proposed machine learning approaches for classifying hemipteran waveforms do not apply well to waveforms from blood-feeding arthropods. This suggests that analyzing species with distinct waveform patterns may require differing machine learning approaches. The findings of this study will facilitate the development of an automated waveform identification tool that researchers can use to score waveforms from mosquitoes and other blood-feeding arthropods.